liquid ai's new on-device agent looks a lot like something apple ordered

the information recently reported that apple is racing to bring powerful ai onto the iphone and is shopping for acquisitions in the on-device inference space. liquid ai – a cambridge-based startup focused on efficient inference – is reportedly on their radar

on the same day the story dropped, liquid ai released lfm2.5-8b-a1b

what it is. a small edge model built for agentic tasks – tool calling, instruction following, complex reasoning – designed to run on consumer hardware with no cloud and no data leaving the device. it has 8b total parameters but only 1b active at inference time

what changed. compared to its predecessor, context window expanded from 32k to 128k tokens, training data tripled from 12t to 38t tokens, and it is now a reasoning model – producing a chain of thought before each answer. hallucination rate dropped significantly

speed. 253 tokens/sec on an apple m5 max, around 30 on a phone. supports llama.cpp, mlx, vllm, and sglang from day one across apple silicon, amd, intel, and nvidia hardware

performance. on tau² telecom – a key agentic benchmark – it scores 88.07, the highest of any model under 10b active params, beating same-size rivals by over 60 points and outperforming models with 4x more active parameters. on math500 it leads gemma-4-e4b-it by nearly 24 points at identical parameter count

whether liquid ai is already working with apple or simply building exactly what apple needs, the timing is hard to ignore